Reconciling Explanations in Multi-Model Systems through Probabilistic Argumentation

Multi-model AI systems create explanation conflicts that single-model approaches cannot solve: The paper identifies a critical gap where explanations from different AI models within the same system can contradict each other, preventing users from understanding the system's overall decision-making process—a problem particularly concerning in high-stakes domains like healthcare and finance.

Probabilistic argumentation can reconcile conflicting explanations: Rather than treating contradictory explanations as failures, the proposed framework uses probabilistic argumentation to map uncertain information from different models into a coherent narrative, helping stakeholders understand how multiple AI components reached their conclusions.

Explanation reconciliation should reflect different user perspectives: The framework incorporates user-defined criteria including optimism, pessimism, and fairness perspectives, recognizing that different stakeholders (doctors, patients, regulators) may prioritize different aspects of an explanation and require tailored approaches to transparency.

Computational efficiency matters for practical XAI deployment: By introducing the "relative independence assumption," the authors address a practical challenge in making their framework scalable—reducing the computational burden of generating explanations so that responsible AI systems remain feasible to deploy in real-world settings.

Holistic explanations should align with how humans think: The framework is explicitly designed to generate explanations that match human cognitive processes rather than simply listing technical factors, suggesting that true transparency requires explanations designed for human understanding, not just technical completeness.